This study focuses on a specific region within the United States – the southeastern area classified by the United States Department of Agriculture (USDA) as Region 8 (Fig. ). This region spans approximately 73 million hectares of forest across 13 states, with a wide variety of forest types and disturbance regimes, making it an ideal area for studying large-scale forest dynamics and disturbances.

The forests in this vast region exhibit high diversity, with a mix of broadleaf and coniferous species flourishing across a spectrum of climate conditions. From the humid subtropical zones of coastal and lowland areas to the temperate mixed forests of higher elevations, these ecosystems support a rich array of ecological functions and forest management practices, each uniquely shaped by the interplay of climate, topography, and species composition. In the Interior Highlands and Interior Plains (the central and northern parts of the study area), low-elevation mountains and flat terrain are dominated by deciduous species, including White Oak (Quercus alba), Red Oak (Quercus rubra), and various Hickory species (Carya spp.). In contrast, the Atlantic Plains (southern regions), with their wetter conditions and extensive wetlands, support deciduous species like Baldcypress (Taxodium distichum), Water Tupelo (Nyssa aquatica), and Sweetgum (Liquidambar styraciflua). The Appalachian Highlands, with higher elevations and cooler climates, contribute to further diversity, featuring mixed species such as Sugar Maple (Acer saccharum), Yellow-poplar (Liriodendron tulipifera), and Eastern White Pine (Pinus strobus). A comprehensive species list is available in Appendix Table  and the US Forest Service Map for Tree Species in the Southeastern United States .

The study region spans a broad climatic gradient typical of the southeastern United States. Mean annual temperatures generally range from approximately 12–14 °C in the northern and higher-elevation areas to 20–22 °C in the southern coastal and lowland regions. Annual precipitation is relatively high across the region, typically ranging from about 1100 to 1600 mm, with higher rainfall in coastal zones and the Appalachian Highlands. The region is characterized by humid subtropical conditions in the south and east, transitioning to more temperate climates at higher elevations and latitudes. In addition, large portions of the study area lie within the North Atlantic hurricane track, making the region frequently exposed to tropical storms and hurricanes that bring strong winds and heavy rainfall, particularly in coastal and lowland areas. These climatic gradients and extreme weather events strongly influence forest composition, productivity, and disturbance regimes, including insect outbreaks, storm damage, and drought-related stress.

This diverse forest composition, with varying ecological zones and management strategies – from conservation-driven national parks to privately managed timberlands – creates a complex landscape in which disturbances manifest in various ways across the region. We do not focus on a single tree species or management practice, but instead conduct a complex yet holistic analysis of disturbances applicable to a wide range of forest types.

To analyze forest disturbances affecting this region, we use three key datasets: the Insect and Disease Survey (IDS), Sentinel-1 Change-Detection (S1CD), and Tree Canopy Cover (TCC). These datasets form the foundation of our analysis, providing complementary information on forest disturbances, structural changes in canopy cover, and the presence/absence of trees. Table  summarizes their key characteristics, with further details discussed in the following sections.

Throughout this paper, we use consistent terminology to refer to spatial representations of disturbances. We use the term polygon to describe the vector-based geometries in the IDS dataset, reflecting the original IDS data format and nomenclature. In contrast, patch refers to the disturbance outlines derived from the S1CD dataset in raster format, which are later converted to polygons for analysis.

The Insect and Disease Survey (IDS), is conducted by the United States Department of Agriculture (USDA) Forest Service and monitors forest damage and mortality resulting from various disturbance agents. Since 1997, the IDS has monitored disturbance agents across the United States and its territories, including fires, wind, bark beetles, defoliating insects, and fungal diseases.

Disturbances are mapped as geo-referenced polygons that represent areas where damage was observed, though not all trees within a polygon are necessarily affected. For each polygon, surveyors record a range of attributes ; those used in this study include the affected tree species, the causal agent, the survey year, the damage type (e.g., mortality, defoliation percentage, crown dieback, branch breakage, or other non-mortality damage), and the severity class, expressed as the percentage of live and damaged trees within the polygon (Very Light: 1 %–3 % through Very Severe: >50%). For the causal agents, the IDS database includes more than 1000 fine-grained disturbance agent codes, which are hierarchically organized into 28 broader categories. To reduce interpreter uncertainty and ensure analytical consistency, the 63 bark beetle and 210 defoliator agents were aggregated to their broader categories in IDS (Table ), while wind, formally a subgroup of the abiotic damage category, was considered separately due to its distinct disturbance characteristics. The resulting aggregated disturbance categories used here, namely wind, bark beetles, and defoliators, are referred to as DCA_ID.

In our study region 8, the IDS recorded a total of 2992 disturbances from 2017 to 2020. Bark beetle outbreaks were the most frequent (1177 events), followed by wind (478) and defoliators (213). Annual disturbance counts varied considerably, from 210 events in 2017 to 1374 in 2018. The total disturbed area also varied by type and year: defoliation affected the most significant area in 2018 (119 km²), while wind caused the largest area loss in 2020 (343 km²). Bark beetle outbreaks were widespread but generally affected smaller areas per event, whereas wind and defoliation events often covered larger contiguous areas.

IDS data is collected primarily through aerial detection surveys (ADS) conducted by trained professionals. While ground-based surveys are used to verify specific disturbances or assess localized events, ADS provides the most effective method for large-scale disturbance monitoring, especially for extensive forest areas .

Although mapping disturbances across large areas, such as the United States, would be impossible without ADS, this method presents several challenges. Surveyors must contend with factors such as the aircraft's high speed (approximately 100 mph), the angle of observation, and the complexity of detecting disturbances in a constantly changing landscape. Despite the aid of geo-referencing tools, these challenges often result in imprecise delineations, sometimes including healthy trees or non-forest areas, especially during large-scale outbreaks when time constraints limit accuracy . This uncertainty is explicitly acknowledged in the IDS protocol: for each mapped polygon, the attribute “percent affected” provides a standardized estimate of the proportion of trees within the area that the identified disturbance agent impacts. Although quick decisions are necessary during ADS, surveyors generally correctly attribute disturbance agents, making false positives (Type I errors, where a disturbance is recorded but did not actually occur) rare. However, compound disturbance events, i.e., events that occur in quick succession, simultaneously, or that trigger one another, can create overlapping effects. For example, a drought may increase bark beetle activity, or storms may be followed by bark beetle outbreaks. These overlapping influences make it difficult to disentangle individual agents, sometimes leading to false negatives (Type II errors, where a real disturbance impact is missed or misattributed), resulting in specific disturbance effects going unnoticed .

Biotic disturbances, such as defoliators and bark beetles, occur within specific temporal windows of activity that must be captured accurately through timely ADS. However, weather, funding, and surveyor availability can delay flights, preventing surveys from being conducted during these critical windows. As a result, disturbances may be missed or monitored at suboptimal times . Additionally, some disturbances may not be detectable until after they occur, resulting in a time lag between the actual disturbance and its detection, as seen with bark beetles . Due to the large areas that must be covered, flight patterns often change from year to year, and some regions may not be surveyed at all in a given season. Survey coverage follows specific flight paths rather than encompassing every forested area, and the extent of the flown area is made available annually as a geodatabase. This information is particularly important for interpreting records from 2020 and 2021, when COVID-19 restrictions limited aerial surveys and protocols shifted toward greater use of remote sensing. This variability in flight coverage can introduce biases and inconsistencies, particularly in disturbance attribution, since areas may not always be flown at the optimal time for detecting damage.

Given these spatial and temporal uncertainties, ground verification is used to validate aerial survey data and investigate anomalies. However, due to time, funding, and data volume constraints, less than 1 % of the mapped polygons are verified annually .

Despite these challenges, the IDS dataset offers invaluable insights into large-scale forest disturbances. With nearly 27 years of data since 1997, it provides a unique perspective for analyzing short- and long-term disturbance patterns. Another strength is its broad spatial extent, spanning the entire United States, including remote and difficult-to-reach areas where ground-based monitoring is impractical. Importantly, it covers both public and private lands, thereby reducing biases related to forest management practices, land ownership, and tree species. The dataset captures a wide range of disturbance agents, including fires, wind, bark beetles, defoliating insects, and fungal diseases, along with detailed metadata on damage severity, affected tree species, and other critical factors. While it has limitations in spatial and temporal accuracy and verification, its extensive coverage, variety of disturbance agents, and detailed metadata make it an indispensable tool for understanding forest dynamics and disturbance patterns across the United States.

To focus exclusively on forested areas, we used the Tree Canopy Cover dataset NLCD TCC; from the Forest Service of the U.S. Department of Agriculture. The data is available for CONUS for the years 2015–2020 at a 30 m × 30 m spatial resolution and used to distinguish forested areas from non-forest areas in the study.

Here, we aim to produce an independent, satellite-based dataset of forest disturbances using Sentinel-1 C-Band Synthetic Aperture Radar (SAR) data . Unlike optical sensors, SAR is unaffected by cloud cover or sunlight, allowing consistent, year-round monitoring at high temporal (12 d) and spatial (20 m × 20 m) resolution . The C-band radar (5.4 GHz, 5.6 cm wavelength), and particularly the Vertical transmit/Horizontal receive (VH) backscatter, is sensitive to canopy water content and structure , making it particularly well-suited to capture the three disturbance types that are the focus of this study. Bark beetle outbreaks reduce canopy moisture , defoliators minimize leaf area and wind can remove branches or uproot entire trees, creating gaps and altering canopy structure – all of which can be detected by radar even when foliage remains green to human observers or spectral satellite sensors. Bark beetle outbreaks reduce canopy moisture , defoliators minimize leaf area , and wind can remove branches or uproot entire trees , creating gaps and altering canopy structure – all of which can be detected by radar even when foliage remains green to human observers or spectral satellite sensors .

Here used the VH data from the Sentinel-1 Backscatter Datacube produced by , extracted for 17 tiles covering our study region for the period 2016–2021 (see Fig. , Table ). These data are orthorectified and topographically corrected, and provided on a 20 m × 20 m resolution grid in the EQUI7 grid system .

We used Planet data in this study to manually validate the new dataset and statistically assess whether the methodology produced a significantly improved product compared to IDS. We acquired the PlanetScope imagery through the Planet Education and Research (E&R) Standard Program , which provides limited, non-commercial access to PlanetScope and RapidEye imagery for university-affiliated students, faculty, and researchers. Access to the program requires a valid university email address; in this case, imagery was obtained using an institutional email from the University of Leipzig, granting permission to use PlanetScope data for research purposes.

PlanetScope imagery consists of high-resolution optical data acquired from a constellation of CubeSats, with a nominal daily revisit, although actual coverage can be affected by clouds, acquisition gaps, or satellite scheduling. Each scene has a spatial resolution of approximately 3–5 m and includes four spectral bands: Blue, Green, Red, and Near-Infrared, suitable for monitoring vegetation and detecting forest disturbances.

RapidEye imagery complements PlanetScope with 5 m multispectral data, including a Red Edge band in addition to Blue, Green, Red, and Near-Infrared, enhancing vegetation health assessment. Both datasets are orthorectified and radiometrically calibrated, allowing consistent analysis over time and across study areas. Their high temporal frequency and spectral information make them particularly useful for tracking fine-scale forest disturbances and vegetation dynamics.

The original IDS dataset provided extensive detail, including 47 metadata variables describing disturbance agents, host tree species, damage severity, disturbance year, tree mortality rates, and additional contextual information. However, the complete list of these details was beyond the scope of our analysis. Following the approach described by , we implemented several preprocessing steps to clean, simplify, and streamline the data, thereby aligning it more closely with our research needs.

We focused on events within the study area (USDA Region 8) and analyzed disturbances from 2016 to 2020. This period aligned with the availability of Sentinel-1 data (launched in 2014, with consistent coverage beginning in 2016). To select the disturbances of interest, we extracted eleven key variables from the IDS metadata. The damage-causing agent (DCA_ID, DCA_CODE) was recorded both as text and as a numerical code, representing agents such as wind, bark beetles, or defoliators. The survey year (SURVEY_YEA) indicated when the disturbance was observed, and the region (REGION_ID) specified the USDA region in which it occurred. The damage type (DAMAGE_TYP, DAMAGE_T_1) was recorded both as a text description (e.g., defoliation, mortality, crown dieback) and as a numerical code. The percentage of affected trees (PERCENT_AF) indicated the proportion of trees within the polygon that were damaged, while the host species (HOST, HOST_CODE) identified the affected tree species in both text and numeric form. Finally, the geometry variable defined the spatial extent of the disturbance polygon. Together, these variables allowed us to identify disturbance agents, quantify the extent of damage, and link disturbances to specific tree species. Based on the 26 disturbance agents identified by IDS, we aggregated disturbances into three categories: wind, bark beetles, and defoliators. Following , we aggregated spatially overlapping polygons from the same survey year and disturbance category into single events (ID_E).

To prioritize the most severe disturbances, which are more likely to be identified by satellite, we retained only events classified as “Severe” or “Very Severe” (more than 30 % of trees affected), as well as polygons without a recorded value for the percentage of affected trees (PERCENT_AF). According to , some IDS polygons contained predominantly non-forest areas, had irregular shapes, or covered unrealistically large extents. However, such cases were extremely rare, with fewer than 1 % of polygons exceeding 15 km². Despite their rarity, these large polygons represented very extensive and poorly defined disturbance areas (mean size ≈121 km²), and we removed them to avoid introducing bias and uncertainty into our analysis. While excluded only events larger than 2000 km², we applied a stricter filter by excluding polygons larger than 15 km². We did not apply a tree canopy cover filter because, as discussed in Sect. , the IDS may contain spatial location inaccuracies, and we did not want to exclude events that were spatially misaligned but still relevant.

Forest disturbances often occur in combination, where one disturbance may trigger or amplify another (e.g., bark beetle outbreaks following drought stress; ). This pattern is often observed in IDS data, where multiple polygons overlap in both space and time. However, given our aim to develop a dataset that could serve as a reference for training forest disturbance classification models, we focused on disturbances that occurred independently to capture a cleaner signal for each disturbance type. We excluded all overlapping events within a ± 5-year window to eliminate spatially and temporally compounded events. We kept overlapping events with the same disturbance type, as these likely represented continuous disturbances.

To restrict our analysis to forested areas, we first processed the Tree Canopy Cover (TCC) dataset for the years 2015–2020. We resampled the original 30 m × 30 m raster to 20 m × 20 m resolution using nearest-neighbor interpolation to ensure consistency with the Sentinel-1 SAR data. We then reprojected the dataset to the EPSG:4326 coordinate reference system and clipped it to the boundaries of the study region.

Pixel values in the TCC dataset ranged from 0 to 255, representing scaled canopy cover percentages (0=0%, 255 ≈ 100 %). We normalized the values to a 0–1 range for further processing. For each year from 2015 to 2020, we used the previous years TCC data to mask the following year’s Sentinel-1 SAR files. We classified pixels with TCC ≥ 30 % as forested, and pixels below this threshold as non-forest. We applied a 30 % TCC threshold, which is stricter than the FAO forest definition (≥ 10 % canopy cover; ), to ensure that detected disturbed patches correspond to forest disturbances and not changes in vegetation structure driven by other processes, such as agricultural practices. Using a 10 % threshold yielded very similar results, with fewer than 80 additional events per disturbance type.

Based on the Sentinel-1 VH backscatter data for the years 2016–2021 over our study region described in Sect. , we applied the pixel-wise change detection method described by to identify forest disturbance patches at 20 m × 20 m spatial resolution. The approach relies on Recurrence Quantification Analysis RQA; of time-series of VH data, which is applied to each pixel individually. The purpose of the RQA approach is to quantify the similarity between time steps in a time series. Here, for each calendar year within the study period (2016–2020), we select a two-year window spanning July of the previous year to June of the following year in order to evaluate the presence of a disturbance in that calendar year (i.e., 60 time-steps). The main elements of the RQA algorithm are illustrated in Fig. .

In brief, the method plots the time series against itself to generate a Recurrence Plot (RP), where each point in the time series is compared with all others (i.e., 60×60 time-steps in this case), and a recurrence condition (true/false) is assigned based on the absolute difference of the backscatter signal. Here, the recurrence threshold (ϵ) was set to 3 dB. We computed the RP following by using the RecurrenceAnalysis.jl package.

By definition, the main diagonal is composed of ones (Fig.  c). Then, a Recurrence Rate (RR) can be calculated for each diagonal of varying length, defined as the fraction of recurrent points along the diagonal relative to its length, resulting in a recurrence line quantifying the recurrence fraction per length (Fig. d). The slope of this line (RQA-TREND) reflects the temporal dynamics of the pixel: a slope of zero indicates no change over the whole period, whereas negative or positive slopes indicate decreasing or increasing recurrence, respectively. We applied an RQA-TREND threshold of −1.28 to distinguish between stable and disturbed pixels. This threshold was selected based on a preliminary analysis of the 5th percentile of the RQA-TREND over a known disturbance event, consistent with the results for deforestation events. The choice of 5th percentile is based on as an appropriate value to detect forest loss in SAR data, and validated in for case studies of deforestation. To test the robustness and validity of this choice on our study area, we performed a sensitivity test on a subset of 3 Planetscope manually labeled disturbance patches (see Sect. ), one for each disturbance (see Fig. ). We tested a range of RQA-TREND thresholds and evaluated their performance using the F1 score and Intersection over Union (IoU) for each disturbance type (DCA_ID). The optimal thresholds differed among disturbance agents, with stronger trends performing better in detecting bark beetles and wind (−2.78), and more moderate thresholds for defoliators (F1 peaking at -1.28 to −0.78, with very similar values).

Given the preliminary sensitivity test, we selected the threshold of −1.28, which lies within the overlapping range of effective thresholds across the three disturbance types. This compromise provides a balanced performance between detecting true disturbances and limiting overestimation, although it is somewhat restrictive for capturing defoliator disturbances. We classified pixels with RQA-TREND values above −1.28 as showing no significant change, and those below this threshold as indicating a structural change in that year. This process produced a binary raster per year, where each pixel indicated whether a structural change in forest cover occurred during that year. Finally, we projected this resulting WGS84 coordinate system (EPSG:4326) to ensure compatibility with the other spatial datasets used in this study.

To exclude non-forest areas from the raster, we used the preprocessed CONUS Tree Canopy Cover (TCC) dataset to filter out non-forest areas (refer to Sect. ). Lastly, we converted the binary raster pixels into polygons to ensure a consistent, polygon-based dataset format. In the following, we refer to this dataset as S1CD (Sentinel-1 Change Detection).

The two datasets – the IDS and the yearly change detection via S1CD – provided complementary and independent information on forest disturbances and their causes. We first compared the two datasets spatially and temporally to identify S1CD-disturbed patches that best matched the reported IDS disturbances. To focus on refining the spatio-temporal attributes of known disturbances, rather than detecting new events, we considered only S1CD elements located within a buffer around IDS polygons, excluding all areas outside both the buffer and the flown survey paths. With this, we aimed to extract a subset of disturbed patches that could be enhanced with disturbance-agent information from IDS. To account for spatial uncertainty in the aerial survey data , we tested buffer distances from 100 to 1000 m, and selected 500 m as the optimal balance between retaining relevant events and minimizing noise. Accordingly, we applied a 500 m buffer around each IDS event. We used a temporal buffer of ± 2 years for the S1CD data to account for the inherent temporal uncertainties in the IDS and S1CD datasets. This choice was necessary because the Sentinel-1 change detection algorithm introduced a ±0.5-year uncertainty due to its time window from July of the previous year to June of the following year (Sect. ). Additionally, reported a ± 3.7-year uncertainty in the disturbance years recorded in the IDS dataset compared to forest inventory data (FIA). Considering these overlapping uncertainties, we opted for a ± 2-year buffer to achieve a more reliable temporal alignment between the two datasets.

We then intersected both the IDS and S1CD datasets (see Fig.  and restricted our analysis to areas along the survey's flight paths where IDS data were available. Since multiple S1CD patches could overlap with a single IDS disturbance, we treated them as scattered patterns from the same event and combined them into a single unified multipolygon. We then removed any aggregated disturbances larger than 15 km², aligning the S1CD-derived dataset (S1DM) with the same size threshold used for IDS, thus focusing our analysis on small- to medium-sized disturbances.

The resulting dataset, Sentinel 1 Disturbance Mapping (S1DM), contained all disturbed patches detected by Sentinel-1 that aligned with IDS-recorded disturbances in space and time, given the uncertainty considered.

We recorded each disturbance event as a separate entry in a GeoDataFrame (Table ). We linked events spanning multiple years based on spatial adjacency and continuity of the affected area, with all related entries assigned the same event ID (IDX_D). Recording each year separately allowed us to track temporal changes in shape and extent, providing insights into the evolution of each event over time.

To assess the spatial agreement between IDS and S1DM, we first analyzed the direct overlap of polygons without applying the 500 m buffer used in previous alignment steps. For each disturbance type (Fig. ), we calculated: (a) the percentage of IDS polygons covered by S1DM, calculated as Percentage=A(IDS∩S1DM)A(IDS)×100, and (b) the percentage of S1DM polygons that overlap with IDS, expressed as Percentage=A(IDS∩S1DM)A(S1DM)×100.

To complement this, we then examined the spatial characteristics of individual S1DM disturbances compared to IDS (without buffer). We calculated three key metrics – the disturbance area, the convex hull area, and the centroid distance – to quantify differences in size, shape, and position. We measured the disturbance area AD as the actual size of each detected patch, providing a direct measure of the affected area. While IDS recorded each disturbance as a single contiguous polygon, S1DM often contained multiple smaller, scattered polygons forming a multipolygon. To account for the overall spatial spread of the fragmented area, we computed the convex hull area ACH, which encloses all individual polygons within the smallest possible convex boundary. This metric helped us understand the overall spatial spread of an event, providing insight into how disturbances extend spatially compared to IDS. Lastly, we calculated the centroid distance Δcentroid, as the Euclidean distance between the centroids (CP=(xC,yC)) of corresponding polygons in the two datasets. These metrics allowed us to quantify differences in size, shape, and spatial positioning between IDS and S1DM disturbances, providing insights into the degree of alignment between the two datasets in terms of scale and location.

We compared the original IDS dataset with our new S1DM dataset to assess whether integrating Sentinel-1 data improved spatial accuracy and boundary delineation for disturbance events. To further validate the precision of S1DM, we used PlanetScope for spatial verification.

We selected 15 events per disturbance type where the disturbed areas could be clearly identified. We used imagery from the PlanetScope and RapidEye satellite constellations (Planet Labs Education and Research Standard Account; accessed August–September 2024 ), captured by the Dove satellites with four spectral bands (blue, green, red, and near-infrared), a nominal daily temporal resolution, and a spatial resolution of 3–5 m. We selected only images with less than 10 % cloud cover to ensure clear visual interpretation. For each event, we visually interpreted and delineated the disturbed areas.

Although we systematically interpreted the imagery, we acknowledge that it may be subject to bias. For example, for some recurrent events, it was difficult to distinguish the original disturbance from subsequent changes, such as regrowth, new tree-line damage, or clear-cuts. To minimize these issues, we focused on extracting the disturbance outline as accurately as possible, emphasizing the largest and clearest extent of the natural disturbance prior to any clear-cut. For each event found in PlanetScope, we recorded the first disturbance date, any subsequent disturbance years, and, when applicable, the year of clear-cutting.

To quantitatively evaluate the spatial agreement between IDS and S1DM disturbance maps and manually delineated reference polygons, we computed two complementary metrics for each event. Manual polygons, visually validated using PlanetScope imagery, were considered the ground truth. For each polygon of IDS or S1DM, we computed the spatial overlap (Eq. ) and the Intersection over Union (IoU) (Eq. ) with the corresponding manual polygon.

The overlap (O) between the polygon areas was defined as the fraction of the manual polygon area covered by the candidate polygon, with values ranging from 0 % to 100 %: 1O=|Acandidate∩Amanual||Amanual|×100. The Intersection over Union (IoU) ranges from 0 (no overlap) to 1 (perfect match) and was calculated as: 2IoU=|Acandidate∩Amanual||Acandidate∪Amanual|, To evaluate the performance of IDS and S1DM in capturing the spatial features of disturbance events, we analyzed the distributions of spatial overlap and Intersection over Union across samples for each disturbance type. We assessed differences between paired observations using a one-sided Wilcoxon signed-rank test, which does not assume normally distributed differences and suits small sample sizes (n=15). The test evaluated whether S1DM systematically captured the spatial characteristics of disturbance events (as identified by manual labeling) better than IDS. We considered a primary significance threshold of α=0.05 (denoted by *) and a secondary threshold of α=0.1 (+). We generated RGB composites from PlanetScope imagery to visually compare the datasets. For each manually labeled event, we selected the best available image after the earliest manually identified disturbance date. We overlaid polygons for manual labels, IDS, and S1DM to illustrate the spatial extent and alignment of the detected disturbance areas.

Specifically, we assessed how many IDS disturbance events contained a corresponding S1DM signal within a 500 m buffer around each IDS patch.

Figures  and  show that the detection effectiveness of S1DM varies by disturbance type. Wind disturbances have the highest agreement, with 98.4 % of IDS wind events having a corresponding S1DM signal within the 500 m buffer (469 of 478 events). Bark beetle disturbances follow, with 76.3 % of IDS patches matched by S1DM (899 of 1177 events). Agreement is lower for defoliator disturbances, with only 67.8 % of IDS-disturbed areas captured by S1DM (144 of 213 events). The percentages in Fig.  refer to totals across all years and match the absolute numbers reported in Fig. .

We analyzed the event size and spatial position of S1DM relative to IDS (without buffer). To quantify this relationship, we examine the spatial agreement between IDS and S1DM for each disturbance type (Fig. ) by calculating (a) the percentage of IDS polygons covered by S1DM and (b) the percentage of S1DM polygons that overlap with IDS. When examining the overlap without a buffer, we find that IDS is covered to a limited extent by S1DM polygons, with median coverage across all disturbance types remaining below 10 %. There is no overlap for most bark beetle events, and for all other disturbance types, IDS is covered by less than 16 % of S1DM polygons in 50 % of the cases. When examining S1DM polygons, the proportion of each polygon's area that overlaps with IDS is generally low for bark beetle and defoliator events. Only for wind disturbances is the median overlap higher, with more than half of the S1DM polygons showing substantial IDS coverage, indicating better agreement between the datasets for this disturbance type.

To explore whether overlap differences arise from variations in polygon sizes, spatial spread, or spatial offsets between IDS and S1DM events, we analyzed these three spatial properties (calculated in Sect. ) for the disturbance events in IDS and S1DM, as shown in Fig.  and Table .

Our findings reveal no notable differences in the individual size of disturbances across different types of events (Fig. a). We find that wind disturbances tend to be the largest events, with 50 % of events having a size of 0.36 km² or smaller (90 % smaller than 5.02 km²), followed by defoliators with a median area of 0.45 km² (90 % below 3.10 km²) and bark beetles with the smallest patches, with a median value of 0.0042 km² (90 % smaller than 0.07 km²).

Overall, disturbance areas derived from S1DM closely match those from IDS, with only minor differences across disturbance types. Bark beetle disturbances show the strongest agreement, with nearly identical median areas (IDS: 0.0042 km²; S1DM: 0.0043 km²) and 90 % of events in both datasets smaller than 0.08 km². For wind disturbances, median S1DM events are slightly larger than IDS (0.38 vs. 0.36 km²), while the 90th percentile areas are nearly identical (5.02 vs. 5.03 km²). Figure a shows a pronounced peak in the probability density function at approximately 4–6 km² for wind-related disturbances, which is consistently observed in both the IDS and S1DM datasets. Defoliator disturbances tend to be larger in S1DM than in IDS, with similar median values (0.50 vs. 0.45 km²) but substantially larger upper-tail events, where the 90th percentile in S1DM is nearly twice that of IDS (6.33 vs. 3.10 km²).

Figure b illustrates the area of the convex hulls enclosing the IDS and S1DM polygons, which serves as an indicator of how spatially dispersed each disturbance event is. In the IDS dataset (without buffer), each event is represented by a single, compact polygon. In contrast, the corresponding S1DM event may consist of multiple scattered polygons (a multipolygon) that encompass all S1CD patches within the 500 m buffer used during S1DM generation. This results in spatially fragmented event structures, as shown in Fig. . Accordingly, Fig. b demonstrates that S1DM events exhibit a substantially larger spatial extent than the corresponding IDS polygons across all disturbance types. These ratios, reflecting the relative spatial spread of S1DM to IDS, are also summarized and calculated from the values in Table . The median convex hull area of S1DM polygons is 5.3 times greater for defoliators (0.49 (IDS) → 2.64 (S1DM) km²) and 5 times greater for wind (0.43 → 2.15 km²) than for the respective IDS polygons. At the 90th percentile, these differences narrow, with the S1DM spread being 4.1 times larger for defoliators (3.68 → 15.27 km²) and 2.2 times larger for wind (5.06 → 11.31 km²). Bark beetle disturbances show the most pronounced increase, with the convex hull area more than 50 times larger for 50 % of events (0.0046 → 0.25 km²) and 17.8 times larger for 90 % of events (0.076 → 1.35 km²), although their overall affected area remains comparatively small, as discussed in Fig. a.

We further analyze the spatial bias between the centers of IDS and S1DM disturbance events (Fig. c). Despite the pronounced differences in spatial extent and fragmentation, the centroids of corresponding events in both datasets remain closely aligned. Most S1DM centroids fall within 200–350 m of their respective IDS polygon centroids, indicating a strong spatial correspondence between mapped disturbance locations. Even at the 90th percentile, centroid distances remain below 1 km for the majority of events, suggesting that while S1DM polygons are more spatially dispersed, their overall positioning relative to IDS remains consistent.

We analyzed the temporal agreement between corresponding events in IDS and S1DM (Fig. ). We assess the temporal agreement between the datasets by identifying, for each IDS event, the S1DM detection year that is closest in time. This allows us to quantify how many of the S1DM-detected disturbances correctly coincide in time with those reported by IDS. Overall, S1DM accurately captures the timing of most disturbance events, as reflected by the high number of events detected at a temporal lag of zero (Lag =0) across all disturbance types. Specifically, for wind disturbances, 443 out of 469 S1DM events (94.5 %) occur in the same year as the corresponding IDS events, indicating excellent temporal agreement. For defoliators, 92 out of 144 S1DM detections (63.9 %) coincide with the IDS year (Lag =0), demonstrating moderate agreement but a clear temporal correspondence between the two datasets. Both wind and defoliator events exhibit only limited temporal mismatch, with detections outside the IDS year (Lag =±1 or ±2) accounting for less than 3 % of wind and 29 % of defoliator events, respectively. This indicates that, for these disturbance types, S1DM tends to identify a disturbance within the correct year. In contrast, bark beetle disturbances display a broader temporal distribution. While the largest proportion of events, 359 out of 899 (39.9 %), are detected by S1DM in the same year as IDS (Lag =0), a substantial number fall within the ±1-year window around the IDS detection year, with 140 events (15.6 %) detected one year earlier and 157 events (17.5 %) one year later. Notably, 212 events (23.6 %) are detected by S1DM approximately two years earlier than IDS, representing the second-largest peak after Lag =0. This pattern suggests that bark beetle disturbances may be identified earlier in radar-based data, potentially capturing the onset of canopy-structure changes before visible defoliation is recorded in IDS.

The temporal agreement of the S1CD signal should, however, be interpreted cautiously, as we find prolonged signals within the ±2-year time window used for the trend-break detection Fig. . For example, due to the detection algorithm, it is possible for a wind event to be detected over multiple consecutive years, although such events are in fact abrupt. Furthermore, temporally continuous signals may reflect other changes, such as clear-cutting or post-disturbance restoration activities. This issue will be further explored in Sect. .

Since both the S1DM and the IDS datasets exhibit spatial and temporal uncertainties (see Sect. ), we assessed their temporal and spatial accuracy against a set of 15 manually labeled events per disturbance type, derived from PlanetScope images (see Methods). Figure  illustrates three representative disturbance events for wind (first row), bark beetles (second row), and defoliators (third row), where the polygons of our manual labels, IDS and S1DM, are overlayed on the RGB composite of the first image showing a disturbance event in that location. For these examples, S1DM appears to capture the spatial extent of disturbances more accurately than IDS. For example, for defoliator and bark beetle events, IDS identifies a smaller area than the RGB composite indicates, whereas for wind disturbances, both IDS and S1DM appear to identify disturbed areas that are not evident in the RGB composite (and thus our manually defined polygon). Overall, the S1DM polygons also appear to show more patchy and irregular spatial outlines compared to both manual and IDS polygons, reflecting the finer resolution and method used for their generation.

In Fig. , we quantify the spatial agreement of IDS and S1DM polygons with the 3 × 15 manually labeled polygons from PlanetScope imagery, using both percentage overlap and the Intersection over Union (see Methods Sect. ). S1DM polygons show substantially higher agreement with manual labels than IDS for wind and bark beetle disturbances, with median values (see Table ) of IoU (panel (a)) of 0.157 and 0.076 for wind and bark beetle events detected by S1DM, respectively, compared to 0.092 and 0.007 for IDS. For the spatial overlap, S1DM estimates a median value of 44.3 %, 20.1 %, and 22.2 % for wind, bark beetles and defoliators, compared to 12.3 %, 0.7 %, and 22.7 % for IDS.

One-sided Wilcoxon rank tests indicate that the improvement in the spatial agreement is significant for both wind and bark beetles with both metrics, although for wind only at 90 % significance for the Intersection over Union. No statistically significant improvement is observed for defoliator disturbances according to the one-sided Wilcoxon rank test. It is worth noting that the Intersection over Union remains generally low, consistently below 0.4, while the percentage overlap ranges from 0 % to 60 %. This might be due to uncertainties in both S1DM and IDS, as well as to the visual labeling approach that relies on the RGB signal.

Here, we produced new forest disturbance maps for the southeastern USA covering the period 2016–2021 based on the Sentinel-1 radar backscatter trend detection algorithm by , S1CD. This algorithm enhanced forest disturbance monitoring by leveraging Sentinel-1's frequent, consistent, and cloud-penetrating radar data to detect significant structural changes in forests. S1CD effectively identified disturbances associated with structural damage, such as bark beetle infestations and storms. As shown in Fig. , S1CD successfully detected over 75 % of IDS-recorded wind and bark beetle disturbances in the study region and selected period. However, detection rates dropped significantly for more subtle disturbances, such as those caused by defoliators, with only 67 % of disturbances reported by IDS identified by S1CD.

This limitation is consistent with known challenges in remote sensing of forest disturbances, where gradual or low-intensity disturbances often remain undetected . In this case, the low sensitivity of S1CD to subtle and slow-onset disturbances may stem from the change detection threshold, which we set at −1.28 based on the method described in and correspondence with the authors. Since the method was optimized to detect abrupt, permanent structural changes, disturbances that unfold gradually or cause only minor canopy loss are likely to have gone undetected. Although lowering the threshold could improve the sensitivity to these subtle disturbances, it would risk increasing noise, leading to more false positives and reducing the detection reliability. Thus, while the Sentinel-1 Change Detection method applied here was highly effective in capturing severe disturbances such as wind-throw and bark beetle outbreaks, its limitations in detecting milder, gradual, or recovering disturbances must be acknowledged.

Additionally, in our IDS preprocessing, we excluded all temporally compound events to retrieve a clear disturbance signal for each disturbance agent. However, by excluding such events, we introduced bias into the dataset, so that our hybrid S1–IDS dataset (S1DM) should not be considered for exhaustive characterization of disturbance occurrence in the study region. While not representative of all disturbance events, the dataset provides a set of “noise-free” disturbance events for each of the three agents considered, which should allow for better training of unsupervised classification models. Below, we discuss the improvements introduced by our hybrid dataset.

In addition to providing an improved hybrid forest disturbance reference dataset for disturbance agent classification, our approach aimed to move beyond the traditional, labor-intensive method of manual labeling. The Sentinel-1 Change Detection (S1CD) algorithm enables analysts to bypass the time-consuming task of manually refining disturbance locations, outlines, and timing by using high-resolution radar-based information on forest structure and change. By employing the trend-break detection method of , we successfully automated much of the re-labeling process, creating an efficient and scalable system that requires a smaller contribution from expert knowledge. This not only reduces the need for manual inspection but also offers a viable alternative to manual labor, minimizing reliance on private, sometimes exploitative labor platforms located in developing countries, as is often the case for deep learning label generation .

Our semi-automatic labeling method was initially developed to address specific challenges in reference dataset quality (see Sect. ). A key advantage of our semi-automated labeling method is its adaptability to different regions, as long as some form of inventory-based disturbance data, e.g., the Canadian Forest Service’s National Forestry Database (NFD), the Database of Forest Disturbances in Europe DFDE, the Dataset of wind Disturbances in European Forests FORWIND, or the Database of European Forest Insect and Disease Disturbances DEFID2, is available. Sentinel-1 data, available globally since 2014, provides a consistent and uninterrupted record of forest dynamics. Unlike optical datasets, such as Sentinel-2, which are susceptible to cloud cover and atmospheric interference, Sentinel-1’s radar data enable reliable time-series analysis across diverse climatic conditions. This robustness makes it particularly valuable for monitoring large-scale and long-term disturbances.

However, Sentinel-1 data is affected by terrain-induced distortions in mountainous regions, which can reduce classification accuracy . Users should exercise caution when applying this method in areas with significant elevation changes, as radar signal distortions can impact data reliability. Despite this limitation, our approach remains broadly applicable, especially in flat or gently sloping landscapes, where radar performance is more stable.

In this study, several methodological decisions had to be made that could potentially influence our results. To combine disturbance information from IDS and S1CD, we used a spatial buffer of 500 m and a temporal buffer of ±2 years centered on the IDS disturbance date. This spatial and temporal buffer was introduced to account for the spatial and temporal uncertainty in both the IDS dataset as well as in the disturbance date detection by S1CD.

Previous work by analyzed the spatial and temporal agreement in reported disturbances for IDS and the Forest Inventory and Analysis for the continental U.S. An analysis of spatial buffers of 800 and 1600 m, showed that increasing buffer size introduced additional uncertainty. The disturbance timing differed by 0.7 years on average, for a buffer size of 800 m. We have tested the effect of using different spatial buffer sizes of 100, 250, 500, and 1000 m on the number of events retained for analysis, and found that buffer sizes of 100–250 m resulted in too few events being selected, while 1000 m did not retain considerably more events. Furthermore, our change detection algorithm is based on recurrence analysis trends over 2-year-long windows, meaning that the difference in the timing of disturbance with IDS could be up to 2 years, if the disturbance happened at the end of the analysis period. Therefore, our choice of 500 m and 2-year windows represents a balance between retaining information and minimizing the risk of introducing additional uncertainties. Future work could explore in greater detail the impact of spatial uncertainty between inventory datasets and satellite observations to refine buffer selection.

Another key methodological limitation concerns the selection of the temporal threshold for classifying structural changes. Like most other disturbance detection approaches , our algorithm requires a change in the temporal structure of the VH signal that is sharp enough to be distinguished from normal variability. In this study, we applied a threshold of −1.28 of the RQA-TREND calculated over a 2-year window centered around each year's summer to distinguish between stable and disturbed pixels. While this value was empirically derived from reference datasets, it remains a heuristic choice and can influence the spatial coherence and apparent patchiness of detected disturbances. We tested the effect of this choice by evaluating the performance of the S1CD in detecting the set of manually labeled disturbance patches from PlanetScope for different values of the RQA-TREND and per disturbance type (Fig. ). The results indicate that the optimal thresholds differ among disturbance agents, with wind and Bark-beetle requiring sharp changes in the temporal structure of the SAR signal, while defoliators are better captured if less strict thresholds are used. The selected threshold of −1.28 falls within the broader range of optimal values across disturbance types and metrics, representing a compromise solution that ensures consistent applicability across all disturbance categories. We explicitly acknowledge this dependency as a limitation of our workflow. The analysis highlights that, in principle, a universal “one-size-fits-all” change-detection threshold does not exist. These results are consistent with and underscore the challenges in monitoring from space disturbance with very different spatial and temporal characteristics, as discussed, e.g., by . We therefore recommend adopting disturbance-specific threshold calibration strategies, as global fixed thresholds are known to produce biased results across ecosystems and disturbance types due to differences in vegetation structure, disturbance dynamics, and sensor response .

A key limitation in disturbance detection workflows is the propagation of errors through sequential processing steps; for example, a misclassification in one dataset can influence the final combined disturbance map. Our framework is explicitly designed to reduce error propagation by integrating multiple sources of disturbance information, including Sentinel-1 change detection. While direct field validation would provide the strongest confirmation of disturbances, this is not possible in our case: the events we analyze occurred 5–9 years ago over a very large region, and many affected areas have likely recovered (for example, defoliator impacts often recover within a year). Previous work has explicitly assessed inconsistencies and uncertainties in IDS. compared IDS with independent inventory data and found good temporal agreement with FIA, suggesting that temporal errors in IDS are generally low. Our results support this finding, as IDS timing largely aligns with S1DM (Fig. ), except that some bark beetle disturbances are detected earlier by S1DM than reported in IDS. However, because FIA lacks sufficiently detailed spatial information, spatial event-level errors cannot be quantified using inventory data alone. To address this limitation, we employed Planet imagery as an independent, spatially explicit proxy for validation. Although this cannot fully replace field observations, it allows us to evaluate the plausibility of disturbances in areas where ground data are no longer available.

In this study, our analysis was constrained to areas covered by the IDS dataset, which provides detailed, spatially explicit information on disturbance agents. While Sentinel-1 offers global coverage and the potential to detect disturbances beyond regions surveyed by aerial or ground-based inventories, the lack of independent reference data limits our ability to confidently attribute detected disturbances to specific causal agents in areas outside IDS coverage. Consequently, the current framework cannot fully evaluate the performance of Sentinel-1-based disturbance detection across unsurveyed regions or in landscapes where IDS data are absent. Future work should address these limitations by integrating Sentinel-1 disturbance detection with high-resolution, independent reference datasets, such as manual delineations from Planet imagery, and by using tree canopy cover products to exclude non-forested areas. Expanding the analysis in this way would allow for a more comprehensive assessment of Sentinel-1’s capabilities and its potential for large-scale, autonomous forest disturbance monitoring.

Our study demonstrates that integrating Sentinel-1 spatio-temporal information can help refine disturbance mapping from inventory datasets. The data processing pipeline is provided alongside this study (see GitHub in Code Availability), so that it can be extended to other regions, as well as different combinations of inventory and satellite-based datasets. We discuss some possible applications and considerations regarding the generalizability of this approach below.

The approach proposed here can be easily applied to other inventory datasets and satellite missions, provided they have overlapping spatial and temporal coverage. The structural information inherent in SAR backscatter data and the high spatial resolution of Sentinel-1 offer opportunities to enrich existing forest disturbance inventories, improve temporal and spatial resolution, and reduce uncertainty in forest loss mapping. For IDS, which is already spatially explicit, S1CD improved the spatial delineation of the disturbed patches, especially for bark beetles and wind (Fig. ). Therefore, our study shows that SAR-based change detection provides a valuable complementary source of information to aerial surveys and field inventories. For point-based inventory datasets, such as National Forest Inventories , S1CD provides a means to expand the spatial information of the data, provided that precise coordinates are available for forest inventory plots. The pipeline can also be adapted to other existing satellite-based forest disturbance datasets, such as the Global Forest Change , or the European Forest Disturbance Atlas . The main bottleneck for extending the application to other regions remains the limited availability of suitable inventory datasets, i.e., with large-scale spatial and temporal coverage and detailed and consistent forest disturbance agent information, ideally spatially explicit and with precise coordinates, and that are publicly available and that follow FAIR principles (findable, accessible, interoperable, and reusable). However, recent efforts to compile information on specific forest disturbances, e.g. might still be useful. With appropriate reference datasets, however, there are no inherent technical limitations that prevent extending the methodology to other regions or larger spatial extents. Therefore, initiatives such as European Union's proposal for a Forest Monitoring Framework , aiming to facilitate access to forest-relevant data, including forest disturbances would allow expanding and testing our approach in other regions.

An open question that remains is whether this approach is transferable to regions with complex topography, such as mountainous areas. Our study region is limited to the southeastern United States, which is relatively flat and was selected because of its high incidence of wind disturbances. The absence of large mountains avoids topographic-induced uncertainties known to affect Sentinel-1 observations . While the framework performs well under these conditions, its application in more mountainous regions would require additional testing. In addition, differences in forest structure and species composition may require region- or forest-type-specific adjustments to parameters, particularly for backscatter dynamics and disturbance signatures. In this context, other satellite-based information sources could replace, or complement the S1CD mapping, for example, vegetation indices reflecting other canopy properties from spectral imagery , or hyperspectral information from new missions such as ENMAP .

Various forest disturbance classification algorithms have been developed to separate different disturbance types . The three disturbances considered here, wind, bark beetles, and defoliators, are particularly hard to disentangle, given their particular spatial and temporal characteristics e.g.. With increasing interest in exploring machine learning (ML) or deep learning (DL) models for forest disturbance classification , having large-scale spatially and temporally consistent data on the different disturbance types is key either as training labels for supervised models, or else as evaluation data for new classification models in general. The S1DM dataset addresses this need, in that it contains a subset of IDS-based disturbed patches for these three disturbance types, with refined spatial delineation (Fig. ) and temporal information, as S1CD tends to detect bark beetle, and to some extent also defoliator, disturbances earlier than they are reported in IDS (Fig. ). This reduced set of disturbed patches is still large enough (469, 899, and 144 samples for wind, bark beetles, and defoliators, respectively, over a region of 1 701 214.39 km²) to support the development of dedicated and data-intensive classification algorithms to separate these different disturbance types.

In this way, the refined dataset provides a foundation for the development of adaptive classification frameworks that can generalize across regions, forest types, and disturbance regimes.